1. The Energy Cost of AI at Scale

Artificial intelligence (AI) underpins a growing range of applications, including smart mobility systems, digital twins, and intelligent infrastructure. Its adoption has accelerated rapidly; a recent Microsoft report indicates that approximately one in six people worldwide now use generative AI tools [1], highlighting the pace at which these technologies are being integrated into mainstream use.

This rapid growth of AI entails substantial costs. The training of large-scale models such as ChatGPT-4 (March 2023) and Gemini (December 2023) required investments of hundreds of millions of dollars. Estimates by Epoch AI further suggest that amortized hardware and energy cost for the final training run of frontier models has increased at an annual rate of approximately 2.4x since 2016 [2]. In response, major technology firms have committed to large-scale infrastructure investments, including nuclear power facilities and multi-gigawatt data centers, amounting to hundreds of billions of US dollars globally [3]. Beyond energy demand, current AI training practices show systematic inefficiencies. Conventional AI training pipelines rely heavily on persistently powered servers and over-provisioned hardware. Although renewable energy sources have been partially integrated, the continuous operation of large data centers remains heavily dependent on carbon-intensive energy sources [4]. One preprint study from Harvard’s T.H. Chan School of Public Health reveals that data centers exceed the US’ carbon density average by 48% [5]. These factors signal significant idle energy consumption, high operational costs, and a growing carbon footprint, indicating a mismatch between prevailing infrastructure design and operational profiles of AI workloads.

2. Serverless Computing as a Sustainability Lever

The limitations of existing architectures motivate a revision of how computing resources are supplied for AI training. Rather than continuously expanding energy-intensive infrastructure to meet growing demand, improving computational efficiency represents a direct pathway to curbing the growth of energy consumption as AI scales. Addressing this critical challenge is at the core of the Green Serverless Computing for Resource-Efficient AI Training project conducted at the Center for Environmental Intelligence (CEI), VinUniversity. 

The project investigates the use of dynamic resource management to better align computing capacity with the temporal and spatial characteristics of AI training workloads. In contrast to fixed, always-on infrastructure, serverless computing enables workloads to be executed only when required, with resources allocated on demand and automatically scaled in response to workload intensity. This execution model reduces reliance on persistent over-provisioning while maintaining the scalability and reliability necessary for training large models.

The project is structured around a clear set of research objectives:

• Develop a highly efficient serverless AI training architecture that reduces computing overhead.
• Minimize energy consumption and improve resource utilization through dynamic scaling.
• Lower operational costs make AI training more affordable for startups, researchers, and enterprises.
• Advance sustainable AI practices, aligning with global efforts to reduce carbon emissions and energy waste.

3. From Architecture to Application:

To demonstrate practical relevance, the project focuses on application domains where sustainability and scalability are critical. These use cases illustrate how serverless computing can translate architectural efficiency into measurable environmental and system-level benefits under real-world conditions.

• Digital Twin Platform: In digital twin applications, which require continuous data ingestion and real-time analytics, serverless computing enables low-latency, scalable deployment of virtual replicas across diverse geographic regions. By dynamically allocating computational resources in response to real-time demand, serverless architectures significantly reduce idle energy consumption. This on-demand provisioning mechanism allows digital twin platforms to adapt efficiently to fluctuating workloads while maintaining performance, facilitating development of city-scale digital twins in a more energy-efficient and environmentally sustainable manner.

Figure 1. City-scale visualization of a serverless edge deployment, showing edge nodes overlaid on a road network and the current user-to-node connectivity patterns, alongside per-node resource status (CPU, memory, and container state).

• Sustainable Electric Vehicle Charging: The project also explores sustainable electric vehicle (EV) charging as a key application domain, where computing efficiency directly influences energy system performance. Integrating serverless frameworks into EV charging stations allows for dynamic scheduling of charging and discharging based on real-time demand and grid load. This adaptive approach helps mitigate grid overload during peak periods, improves the utilization of available energy resources, and reduces reliance on fossil fuel-based power generation.

4. Toward Environmentally Responsible AI

As AI adoption accelerates, large-scale infrastructure calls for sustainability considerations. The Green Serverless Computing project demonstrates how architectural innovation, when paired with rigorous research and real-world validation, can significantly reduce the environmental impact of AI training. By linking cutting-edge computing paradigms with pressing sustainability goals, this initiative exemplifies VinUniversity’s commitment to translating research into meaningful societal impact by advancing not only smarter systems but also a greener digital future.

[1] Microsoft, “Global AI Adoption in 2025,” 2026. [Online]. Available: https://www.microsoft.com/en-us/corporate-responsibility/topics/AI-Economy-Institute/reports/Global-AI-Adoption-2025/

[2] B. Cottier et al., “How much does it cost to train frontier AI models?” Epoch AI, 2024. [Online]. Available: https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models

[3] EnergyConnects, “Meta signs multi-gigawatt nuclear deals for AI data centers,” 2026. [Online]. Available: https://www.energyconnects.com/news/utilities/2026/january/meta-signs-multi-gigawatt-nuclear-deals-for-ai-data-centers/

[4] C. Metz, “We did the math on AI’s energy footprint. Here’s the story you haven’t heard,” MIT Technology Review, May 2025. [Online]. Available: https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

[5] G. Guidi et al., “Environmental Burden of United States Data Centers in the Artificial Intelligence Era,” arXiv, 2024. [Online]. Available: https://arxiv.org/html/2411.09786v1